The company says it is applying the approach to the Shopify App, which has more than 300 screens, after rebuilding and publishing Shop in 12 weeks.
The notable design choice is not simply the use of coding models. Helix decomposes a target screen into small checkpoints and blocks progress until each one passes behavior tests, a screenshot-based UI review, two context-isolated adversarial code reviews, and—by default—an engineer sign-off. Each accepted checkpoint becomes a commit; feedback is retained for later work, while teams may selectively enable autonomous runs or parallel migrations.
Shopify says GPT orchestrates matching-state screenshots and Gemini acts as a visual reviewer, while agents work from the existing React Native implementation rather than a large generated specification. This is a concrete example of software-agent workflows being engineered around iterative verification and narrow reviewable changes, not a single generated patch. The post is an account of Shopify's internal system; it does not provide independent measurements of quality, cost, model performance, or general applicability. Source: Shopify Engineering